WLM Neural Mechanics Models

• WLM Neural Mechanics Models are computational frameworks that blend neural networks with Lagrangian dynamics to simulate complex population-level mechanics.
• They integrate second-order dynamics, neuromechanical oscillations, and graph-based relational reasoning to capture both physical and biological interactions.
• The approach leverages techniques like symplectic integration, self-attention for parameterization, and contrastive losses to learn non-conservative and oscillatory behaviors.

The WLM Neural Mechanics Model refers to a family of computational and theoretical models for learning, simulating, and probing the principles of physical and biological mechanics using neural network architectures and dynamical systems. These models span multiple paradigms, including data-driven second-order population dynamics, neuromechanical oscillation in biological organisms, and graph-based relational world models for mechanical reasoning. The following sections synthesize the structure, mathematical formalism, training methodology, and key findings related to WLM Neural Mechanics Models across these domains.

1. Foundations: Neural Mechanics and Wasserstein Lagrangian Dynamics

The WLM Neural Mechanics framework is formalized as learning the underlying mechanics of populations—or agent ensembles—directly from temporal marginals without a prescribed dynamical law. The core principle is to treat the evolution of the system as minimizing a population-level action, specified by a damped Wasserstein Lagrangian: $$\mathcal{L}[\rho(t, \cdot), v(t, \cdot), t] = e^{\gamma t} \left( \mathcal{K}[\rho, v] - \mathcal{U}[\rho] \right)$$ where $\rho$ is the probability density of population states, $v$ the associated velocity field, $\mathcal{K}$ the kinetic energy in $W_2$, $\mathcal{U}$ a general potential energy, and $\gamma$ is a friction parameter. This formulation yields second-order Newton-type equations and permits capturing oscillatory, vortex-like, and non-gradient-flow behaviors that traditional first-order Wasserstein gradient flows cannot (Guan et al., 8 May 2026).

The Hamiltonian equations derived from this Lagrangian are: $$\dot x_t = v_t, \quad v_t = \nabla s_t \qquad \dot v_t = -\nabla \left( \frac{\delta \mathcal{U}}{\delta \rho} \right)(x_t) - \gamma v_t$$ along with the associated continuity equation for the density evolution.

2. Neural Parameterization and Training Protocols

The population-level potential $\mathcal{U}[\rho]$ is approximated by a permutation-invariant network $\Psi_\theta$ acting on empirical measures. For $\rho$0-particle samples $\rho$1, $\rho$2 is realized using multi-head self-attention and multilayer perceptrons: $\rho$3 Acceleration for each particle is obtained as: $\rho$4 The dynamical system is integrated using symplectic Verlet (leapfrog) discretization. For supervised learning, observed marginal point clouds $\rho$5 provide initial positions and (possibly estimated) velocities; training minimizes divergence metrics (Sinkhorn, $\rho$6) between rollouts $\rho$7 and true $\rho$8, with gradients propagated through $\rho$9. A schematic algorithm is:

Step	Action	Objective
1. Initialization	Randomize $v$0 parameters, set friction $v$1	Prepare network for dynamical rollout
2. Rollout	Simulate $v$2 steps via Verlet+network	Generate predicted marginals
3. Loss	Compute $v$3	Minimize distributional error
4. Update	Optimize $v$4 using Adam	Learn population mechanics

This approach enables learning of second-order, non-gradient-flow mechanics from data—without hand-specifying the action functional or deterministic path-law (Guan et al., 8 May 2026).

3. Neuromechanical Principles: Biological Oscillator Models

The Worm Locomotion Model (WLM) expresses neuromechanical gait adaptation in C. elegans as an integrated network of neural, muscular, and body mechanical modules. Each module couples via:

• Mechanical interactions through a viscoelastic beam model,
• Short-range proprioception reflecting feedback of local and anterior curvature,
• Gap-junctional coupling between homologous motor neurons.

The dynamical equations couple neural (bistable oscillator), muscle (activation and contractile force), and body mechanics (viscoelastic beam in fluid). The reduction to a weakly coupled oscillator phase model allows derivation of spatial wavelength $v$5 as a function of phase lag $v$6 between modules: $v$7 Competition between mechanical, proprioceptive, and gap-junction couplings determines wave patterns and their viscosity dependence (Johnson et al., 2020).

Coupling Mode	Dominant Effect	Resulting Wavelength Ratio ($v$8)
Gap-junction only	Synchronous (standing wave)	$v$9
Proprioceptive only	Traveling wave ($\mathcal{K}$0)	$\mathcal{K}$1
Mechanical only	Antiphase ($\mathcal{K}$2)	$\mathcal{K}$3
Full WLM	Weighted sum	Data-matched $\mathcal{K}$4 trend

Viscosity modulates the dominant pathway: at low viscosity, long wavelengths emerge from neural coupling; at high viscosity, mechanical coupling compresses the wavelength, as observed experimentally.

4. Graphical and Relational World Models for Mechanical Reasoning

In the context of symbolic mechanical reasoning (e.g., pulleys), WLM Neural Mechanics Models are proposed as graph-structured encodings combined with relational neural simulators. The architecture entails:

• Diagram Encoder: Parses diagram code (such as TikZ) into a graph, embedding nodes (pulleys, ropes, weights) and edges (physical connections).
• Relational Simulator: Message-passing Graph Neural Network propagates physical quantities (e.g., tension) across the graph for multiple steps, simulating force distribution.
• Language-Fusion Head: Maps simulator outputs to token probability space for regression/classification of mechanical quantities, such as mechanical advantage ($\mathcal{K}$5).

The training protocol combines regression to known $\mathcal{K}$6 values, forced-choice functional discrimination, and contrastive/mirror-consistency losses. Performance metrics include Pearson $\mathcal{K}$7 for regression, $\mathcal{K}$8 for classification, and structured perturbation probes to test generalization and reliance on causal physical features (Robertson et al., 21 Jul 2025).

Quantitative findings indicate that contemporary LLMs using these architectures achieve moderate correlation on $\mathcal{K}$9 estimation ($W_2$0; best models), high $W_2$1 for distinguishing jumbled vs. functional systems ($W_2$2), but degrade to random guessing for deeper connectivity understanding ($W_2$3). Models prominently rely on simple heuristics (counting pulleys) rather than genuine simulation of physical law.

5. Expressivity, Limitations, and Extension Prospects

The principal expressivity of WLM Neural Mechanics lies in its second-order Lagrangian formulation, permitting representation of inertial, damped, and oscillatory population-level dynamics beyond traditional (gradient-flow) methods. This allows modeling of phenomena such as vortex migration, collective flocking, and nonconservative biological motion (Guan et al., 8 May 2026).

Identified limitations include simulation-based training cost for large agent ensembles, lack of path-law identifiability (multiple stochastic processes may produce the same marginal evolution), and absence of stochasticity (deterministic mechanics only). For graph-based relational world models, brittleness and reliance on superficial geometric heuristics are prevalent; models fail to reason robustly about nonfunctional-but-connected systems, indicating a lack of high-fidelity internal simulation (Robertson et al., 21 Jul 2025).

Future improvements include extension to stochastic Hamiltonian dynamics, learning time-dependent potentials, integrating higher-order symplectic integrators, regularization via consistency and visualization probes, and embedding explicit physics constraints directly into the message-passing updates.

6. Significance and Applications

WLM Neural Mechanics Models constitute the first algorithmic framework for inferring population-level mechanics directly from observed distributions, successfully bridging classical mechanics, quantum (via the Lagrangian formalism), and dissipative gradient flows (Guan et al., 8 May 2026). They are empirically validated on synthetic (ocean vortex, Boids) and biological systems (single-cell RNA, embryonic morphodynamics) for both interpolation and extrapolation.

The neuromechanical instantiation in C. elegans provides a concrete mechanistic account of gait adaptation through empirical and theoretical alignment with parameterized neural, muscular, and mechanical interactions (Johnson et al., 2020).

In artificial intelligence and cognition, graphical WLM architectures elucidate the limits and strengths of current language-based models' ability to encode, manipulate, and reason over structured physical world models. This approach anchors the use of output-layer probability distributions as a probe into latent world-model representations and their susceptibility to superficial cues (Robertson et al., 21 Jul 2025).

Domain	Model Role	Empirical Task(s)
Population mechanics	Lagrangian learner	Density forecasting